AI is a sector that is often perceived as having both positive and negative impacts on the environment. On the one hand, AI is able to provide innovative and scalable solutions to tackle the root causes of climate change. On the other hand, AI itself has a high carbon footprint. A University of Massachusetts Amherst study estimates that training a single Natural Language Processing (NLP) model can generate close to 300,000 kg of carbon emissions. According to an MIT Technology Review article, this is “nearly five times the lifetime emissions of the average American car (and that includes the manufacture of the car itself)”.
Therefore, it is clear that environmental sustainability is a key factor to consider when developing AI solutions. Organisations need to pay attention to metrics such as carbon emission and electricity consumption when developing and deploying their models. As the leader in AI optimisation, TurinTech pays special importance to AI’s environmental sustainability. In this article, we discuss how evoML can help you build energy-efficient AI with code optimisation and green metrics.
Build sustainable AI by including green metrics into the process
evoML is an AI optimisation platform that enables businesses to build high performing ML models at speed, with minimum effort. One of the key functionalities of the platform is building machine learning models based on multiple criteria that the users determine. We call this Multi-objective Optimisation. (A detailed description of the evoML model building process is available here.)
At the initiation of the model-building process, users can choose to include green metrics (see Figure 1 –carbon emissions and electricity consumption– for models developed by evoML1. In addition, evoML’s multi-objective optimisation feature enables users to select green metrics as loss functions, along with other performance metrics such as F1, precision, or accuracy.
Figure 1: Option to include green metrics in model building
Once users define preferred criteria, evoML builds and optimises the best model through an iterative process inspired by Darwin’s theory of evolution. evoML determines the winning models and uses their “genes” to create the next generation of models. This evolution process continues indefinitely until the optimal model meets the custom criteria.
Along with a range of performance metrics, evoML provides preliminary green metric values for the best model in a straightforward format (see Figure 2), allowing the user to gauge the environmental impact of the model they are about to deploy.
Figure 2: evoML performance metrics, including green metrics
Achieve higher energy efficiency with code optimisation
Thinking green is just the first step. AI models need to be optimised at the code level to achieve higher energy efficiency. Powered by TurinTech’s proprietary research, evoML is the only platform embedded with Code Optimisation. By improving inefficiencies in code, users can significantly increase their model’s efficiency by over 50% without compromising accuracy and other metrics.
Efficient AI not only reduces carbon emissions but can also lead to better performance at the production level. Imagine being able to gain quicker insights to secure profitable business opportunities or having faster running apps on your phone without draining the battery. By improving the energy efficiency of your model and reducing latency, evoML helps you boost profits and elevate customer experience.
Use evoML visualisations to make well-informed decisions
In addition to preliminary metrics, evoML also provides additional visualisations to help users make more informed decisions about choosing which model to deploy. Figure 3 gives the Pareto efficiency graph that evoML has generated to allow users to choose a model that has both high-performance metrics and lower environmental impact. Generally, a Pareto front gives the set of solutions that have reached the optimal level of selected criteria. In this example, the criteria are precision and negative prediction of carbon emissions2.
Figure 3: Pareto front for selected criteria
Each of the points (red and green points) on the graph in Figure 3 represents a model that has been developed by evoML. Each point has been mapped based on the precision score and carbon emission metric. The green dots represent the optimal models, that is, models that have attained maximum precision while also having minimum carbon emissions. For example, the point labelled ‘A’ on the graph denotes the model with the highest precision but compared to other models its carbon emission is high. Models represented by green points that are further to the right of the graph have lower carbon emissions but are slightly lower in precision. Based on the real-world task the user wants to achieve, the user may choose to deploy a model with lower environmental impact, even at a lower level of precision.
Whatever the final decision may be, evoML metrics and visualisations allow users to make better-informed choices about model development, evaluation, and deployment.
1 CodeCarbon library has been integrated to evoML to obtain carbon emission metrics.
2 The x-axis is calibrated on the negative value of emissions instead of positive values to better visualise the Pareto front.